The questions arrive at the drafting stage now, before anyone files a claim: how often a software agent should stop and check with the person it works for in the middle of a negotiation, and how suspicious it should be of the agent sitting across the table. Anthropic has been putting those questions to law firms, according to two people familiar with the discussions, as it prepares agents that bargain on behalf of users over prices, service disputes and pay.
The Financial Times reported on October 18 that the company has brought lawyers into product design to work out where responsibility sits when an agent negotiates and signs on a customer’s behalf. Among the open items, per those accounts: the point at which a bot must return to its user for approval, and how an agent should treat claims made by another agent without becoming, in the paper’s phrasing, paranoid or prone to conspiratorial thinking.
One proposal under discussion would give agents a signed document from the user that states the instructions and the limits the agent must observe. A second agent could verify that mandate without reading the first one’s prompt, which is the closest thing to a credit check that exists in a market where both sides are software.
That idea addresses a specific failure. An agent asked to prove its authority can currently only describe itself, and a description is not evidence. In research published in April, Anthropic noted that policy and legal frameworks for machine-to-machine bargaining do not yet exist. Work by Michelle Vaccaro at Harvard, who ran more than 180,000 agent-to-agent negotiations, produced a case in which one agent used a prompt injection to try to extract private information from the other. Defending against that requires a rule about what an agent may believe.
The company is not starting from nothing on the legal side. On April 23 it announced a multi-year agreement with Freshfields, the London-founded firm, under which Claude was deployed to 5,700 employees across 33 offices, with adoption rising about 500 percent in the first six weeks. The two organizations said they would co-develop workflows for contract review, drafting and due diligence, and design agentic processes for multi-step legal tasks.
That arrangement gives Anthropic something a customer reference cannot: a firm whose partners argue about liability for a living, embedded in the product cycle while the product is still being defined. Contract review is a task. Deciding who pays when a review is wrong is a different kind of problem.
The liability question has already surfaced in public. In September, Reuters reported that Anthropic had acknowledged its liability limits might not cover claims arising from the actions of autonomous agents. The company’s consumer terms cap its exposure at the greater of six months of payments or 100 dollars, a figure written for a chatbot that answers questions and not for software that commits a user to a price.
Insurers and general counsel have been working the same problem from the other direction. An enterprise that deploys an agent to negotiate with suppliers needs to know whether a bad outcome is a product defect, a user error or a counterparty’s fraud, because each of those lands on a different policy. The contract language that answers it is being written now, in drafts that will be quoted for a decade.
Analysts who follow enterprise software said the company that defines the vocabulary gains an advantage that benchmarks do not confer. If signed mandates become the way agents establish authority, the protocol carries the assumptions of the company that wrote it, including its view of how much autonomy a user should be allowed to delegate.
The distributional question is harder. The Financial Times noted a concern that if the most capable negotiating agents sit behind the most expensive subscriptions, machine bargaining could reinforce existing advantages and widen gaps in bargaining power. A consumer with a basic tier would send an agent with narrower authority and less skill against a supplier running something better, and the outcome would look like a market result.
There is a technical boundary that no amount of drafting resolves. An agent that can be instructed, verified and audited is still making judgment calls about what a clause means, and a signed mandate sets the limits of authority without improving the quality of the interpretation. The document says what the agent may agree to. It says nothing about whether the agent understood the offer.
What the discussions do settle, if they settle anything, is where the rules will come from. Not from regulators, who have moved slowly on agentic systems, and not from standards bodies, whose protocols for machine transactions are still early. They will come from the contracts that vendors and their largest customers sign, and from the first few disputes that test them.
For Anthropic, that is a familiar position and a profitable one. The company has built its commercial identity on being the vendor that enterprise legal departments can defend to a board. Writing the negotiation rules with law firms is the same strategy applied to a product that signs things.
The agents will arrive whether or not the drafting is finished. What remains open is the sentence that will appear in the terms of service, and which party it protects when a deal goes wrong.
